Abstract
In this paper, we study the error behavior of the nonequispaced fast Fourier transform (NFFT). This approximate algorithm is mainly based on the convenient choice of a compactly supported window function. So far, various window functions have been used and new window functions have recently been proposed. We present novel error estimates for NFFT with compactly supported, continuous window functions and derive rules for convenient choice from the parameters involved in NFFT. The error constant of a window function depends mainly on the oversampling factor and the truncation parameter.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Since the restriction to equispaced data is an essential drawback in several applications of discrete Fourier transform, one has developed fast algorithms for nonequispaced data, the so-called nonequispaced fast Fourier transform (NFFT), see [6, 8, 11, 12, 20, 24] and [18, Chapter 7].
In this paper we investigate error estimates for the NFFT, where we restrict ourselves to an approximation by translates of a previously selected, continuous window function with compact support. Other approaches based on the fast multipole method and on the low-rank approximation were presented in [9, 22].
After the seminal paper [24], the similarities of the window-based algorithms for NFFT became clear. In the following, we give an overview of the window-based NFFT used so far. In this construction of NFFT, a window function is applied together with its Fourier transform. This connection is very important in order to deduce error estimates for the NFFT. This made it possible to determine convenient parameters of the involved window function. To develop an NFFT, the necessary Fourier coefficients of the periodized window function can simply be calculated by a convenient quadrature rule. A challenge are error estimations in order to determine the parameters of the window function involved. By \(C({\mathbb {T}})\) we denote the Banach space of all 1-periodic, continuous functions, where \({\mathbb {T}} = {{\mathbb {R}}}/{{\mathbb {Z}}}\) is the torus.
The considered window functions depend on some parameters. Assume that \(N \in 2 {\mathbb {N}}\) is the order of the given 1-periodic trigonometric polynomial which values will be computed by NFFT. Let \(\sigma > 1\) be an oversampling factor such that \(N_1:= \sigma N \in 2 \,{\mathbb {N}}\). For fixed truncation parameter \(m \in {{\mathbb {N}}}\setminus \{1\}\) with \(2m \ll N_1\), we denote by \(\Phi _{m,N_1}\) the set of all window functions \(\varphi :\, {\mathbb {R}} \rightarrow [0,\,1]\) with the following properties:
Each window function \(\varphi \) is even, has a compact support \([-m/N_1,\,m/N_1]\), and is continuous on \({\mathbb {R}}\).
Each restricted window function \(\varphi |_{ [0,\, m/N_1]}\) is decreasing with \(\varphi (0) = 1\) and \(\varphi (m/N_1) = 0\).
For each window function \(\varphi \), the Fourier transform
$$\begin{aligned} {{\hat{\varphi }}}(v) := \int _{{\mathbb {R}}} \varphi (t)\, {\mathrm e}^{-2\pi {\mathrm i}\,v t}\, {\mathrm d}t = 2\, \int _0^{m/N_1}\varphi (t)\, \cos (2\pi \,v t)\, {\mathrm d}t \end{aligned}$$is positive for all \(v\in [-N/2,\,N/2]\).
Note that for fixed \(N_1\) the truncation parameter m determines the size of the support of \(\varphi \in \Phi _{m,N_1}\). If a window function \(\varphi \in \Phi _{m,N_1}\) has the form
with \(\beta > 0\) and convenient functions \(\varphi _1\), \(\varphi _2\), then \(\beta \) is a so-called shape parameter of \(\varphi \). Examples of window functions of the set \(\Phi _{m,N_1}\) are the B-spline window function (28), the modified B-spline window function (30), the algebraic window function (31), the Bessel window function (39), the \(\sinh \)-type window function (45) and the related window functions (see Sect. 5.5) as well as the modified \(\cosh \)-type window function (46) and the related window functions (see Sect. 5.6). Note that the Kaiser–Bessel window function (see [18, p. 393]) and the Gaussian window function (see [18, p. 390]) are not contained in \(\Phi _{m,N_1}\), since these window functions are supported on whole \({\mathbb {R}}\).
The aim of this paper is a systematic approach to uniform error estimates for NFFT, where a compactly supported, continuous window function \(\varphi \in \Phi _{m,N_1}\) is used. We introduce the \(C({\mathbb {T}})\)-error constant
where \(I_N\) denotes the index set \(\{-N/2,\,1-N/2,\,\ldots ,\,N/2-1\}\). As shown in Lemma 2, the uniform error of the NFFT with nonequispaced spatial data and equispaced frequencies can be estimated by \(e_{\sigma }(\varphi )\). Analogously in Lemma 4, the error of the NFFT with nonequispaced frequencies and equispaced spatial data is estimated by \(e_{\sigma }(\varphi )\) too. Therefore in the following, we study mainly the behavior of the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi )\). Our main result is Theorem 2, where we describe a general concept for the construction of a convenient upper bound for the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi )\) with a window function \(\varphi \in \Phi _{m,N_1}\). Applying Theorem 2, we obtain upper bounds for \(e_{\sigma }(\varphi )\) with special window function \(\varphi \in \Phi _{m,N_1}\). We show that the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi )\) of a window function \(\varphi \in \Phi _{m,N_1}\) depends mainly on the oversampling factor \(\sigma > 1\) and the truncation parameter \(m \in {\mathbb {N}} {\setminus } \{1\}\). Since we are interested in NFFT with relatively low computational cost, the oversampling factor \(\sigma \in \big [\frac{5}{4},\,2\big ]\) and the truncation parameter \(m \in \{2,\,3,\, \ldots ,6\}\) are restricted. These parameters \(\sigma \) and m determine the shape parameter \(\beta \) of the window function. For the Bessel window function (39), the \(\sinh \)-type window function (45), and the modified \(\cosh \)-type window function (46), a good choice is the shape parameter \(\beta = 2\pi m\, \big (1 - \frac{1}{2 \sigma }\big )\).
In connection with NFFT, B-spline window functions were first investigated in [6]. In the important application of particle simulation (see [7]), the B-spline window function was also used. Later it became clear that these methods can be interpreted as a special case of the fast summation method, see [15, 19] and the references therein. Based on this unified approach, one can use all the other window functions for this application too. The convenient choice of the shape parameter is of special importance, as shown in [14] for the root mean square error of the NFFT.
In this paper we suggest four new continuous, compactly supported window functions for the NFFT, namely the algebraic, Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window function. The algebraic window function is very much related to the B-spline window function, but much simpler to compute. We show that the Bessel window function (39), \(\sinh \)-type window function (45), and modified \(\cosh \)-type window function (46) are very convenient for NFFT, since they possess very small \(C({\mathbb {T}})\)-error constants with exponential decay with respect to m. It is difficult to design a window function \(\varphi \in \Phi _{m,N_1}\) with minimal \(C({\mathbb {T}})\)-error constant. We prove that the best error behavior has the modified \(\cosh \)-type window function (46) with the shape parameter \(\beta = 2 \pi m \big (1 - \frac{1}{2\sigma }\big )\), \(\sigma \ge \frac{5}{4}\). Further we compare several window functions with respect to the corresponding \(C({\mathbb {T}})\)-error constants for the NFFT. Based on the error estimates of the \(\sinh \)-type window function, we are able to extend the error estimates, see [21], to window functions where an analytical expression of its Fourier transform is unknown, see [3,4,5].
We prefer the use of compactly supported, continuous window functions \(\varphi \in \Phi _{m,N_1}\) by following reasons:
As explained in Remark 1, the NFFT with a window function \(\varphi \in \Phi _{m,N_1}\) is simpler than the NFFT with a window function supported on whole \({\mathbb {R}}\).
The window functions \(\varphi \in \Phi _{m,N_1}\) presented in Section 5 (and their Fourier transforms) have simple explicit forms and they are convenient as window functions for NFFT.
Few window functions \(\varphi \in \Phi _{m,N_1}\) (such as Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window function) possess low \(C({\mathbb {T}})\)-error constants with exponential decay with respect to m. Up to now, the best error behavior has the modified \(\cosh \)-type window function.
The outline of the paper is as follows. In Sect. 2 we introduce the basic definitions and develop the error estimates for an NFFT with a general compactly supported, continuous window function. Important tools for the estimation of the Fourier transforms of window functions are developed in Sect. 3. In Sect. 4 we present a modified Paley–Wiener Theorem which characterizes the behavior of Fourier transforms of compactly supported functions lying in a special Sobolev space. The main results of this paper are contained in Sect. 5. Using the uniform norm, we present explicit error estimates for the (modified) B-spline, algebraic, Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window functions. Further we show numerical tests so that the \(C({\mathbb {T}})\)-error constants of the different window functions can be easily compared.
2 Convenient window functions for NFFT
Let \({\tilde{\varphi }}:\, {\mathbb {T}} \rightarrow [0,\,1]\) be the 1-periodization of \(\varphi \), i.e.,
Note that for each \(x\in {\mathbb {R}}\) the series (1) has at most one nonzero term. Then the Fourier coefficients of \({\tilde{\varphi }}\) read as follows
By the properties of the window function \(\varphi \in \Phi _{m,N_1}\), the 1-periodic function \({\tilde{\varphi }}\) is continuous on \(\mathbb T\) and of bounded variation over \(\big [- \frac{1}{2},\, \frac{1}{2}\big ]\). Then from the Convergence Theorem of Dirichlet–Jordan (see [27, Vol. 1, pp. 57–58]), it follows that \({\tilde{\varphi }}\) possesses the uniformly convergent Fourier expansion
Lemma 1
For given \(\varphi \in \Phi _{m,N_1}\), the series
is convergent for each \(x\in {\mathbb {R}}\) and has the sum
which coincides with the rectangular rule of the integral
Proof
From (2) it follows that for all \(n \in {\mathbb {Z}}\) and \(x \in {\mathbb {R}}\) it holds
Replacing x by \(x + \frac{\ell }{N_1}\) with \(\ell = 0,\ldots , N_1 -1\), we obtain
Summing the above formulas for \(\ell = 0,\ldots ,N_1 - 1\) and using the known relation
we conclude that
This completes the proof. \(\square \)
2.1 NFFT with noneqispaced spatial data and equispaced frequencies
Let \(M\in 2 {\mathbb {N}}\) and \(I_M := \{-M/2, 1- M/2,\ldots ,M/2-1\}\) be given. The NFFT with nonequispaced spatial data and equispaced frequencies is an approximate, fast algorithm which computes approximately the values \(f(x_j)\), \(j \in I_M\), of a 1-periodic trigonometric polynomial
at M nonequispaced nodes \(x_j \in [-\frac{1}{2},\,\frac{1}{2})\), \(j \in I_M\). Using a window function \(\varphi \in \Phi _{m,N_1}\), the trigonometric polynomial f is approximated by the 1-periodic function
with conveniently chosen coefficients \(g_{\ell }\in {{\mathbb {C}}}\). The computation of the values \(s(x_j)\), \(j\in I_M\), which approximate \(f(x_j)\) is very easy. Since \(\varphi \) is compactly supported and thus \({{\tilde{\varphi }}}\) is well-localized, each value \(s(x_j)\), \(j\in I_M\), is equal to a sum of few nonzero terms.
The coefficients \(g_{\ell }\) can be determined by discrete Fourier transform (DFT) as follows. The 1-periodic function s possesses the Fourier expansion
with the Fourier coefficients
where
In other words, the vector \(({{\hat{g}}}_k )_{k \in I_{N_1}}\) is the DFT of length \(N_1\) of coefficient vector \((g_{\ell })_{\ell \in I_{N_1}}\) such that
In order to approximate f by s, we set
Note that the values \({{\hat{g}}}_k\), \(k \in I_N\), can be used in an efficient way. Even if (3) is only known at finitely many equispaced points of \([0,\,1]\), the Fourier coefficients \(c_k(f)\), \(k \in I_N\), can be approximately determined by a fast Fourier transform (FFT). For many window functions \(\varphi \in \Phi _{m,N_1}\), the Fourier coefficients \(c_k({{\tilde{\varphi }}})\), \(k \in I_N\), are explicitly known.
Then for all \(r \in {\mathbb {Z}}\), it holds
In particular, we see that \(c_n(s) = c_n(f)\) for all \(n \in I_N\) and \(c_n(s) = 0\) for all \(n \in I_{N_1} {\setminus } I_N\). Substituting \(k = n + r N_1\) with \(n\in I_{N_1}\) and \(r \in {{\mathbb {Z}}}\), we obtain
Let \(A({\mathbb {T}})\) be the Wiener algebra of all 1-periodic functions \(g\in L_1({\mathbb {T}})\) with the property
Then
is the norm of \(A({\mathbb {T}})\). Obviously, we have \(A({\mathbb {T}}) \subset C({\mathbb {T}})\), where \(C({\mathbb {T}})\) denotes the Banach space of all 1-periodic, continuous functions with the uniform norm
Since \(x_j \in [-\frac{1}{2},\,\frac{1}{2})\), \(j\in I_M\), are arbitrary nodes, we have
Therefore we measure the error of NFFT \(\Vert s - f\Vert _{C({\mathbb {T}})}\) in the uniform norm. As norm of the 1-periodic trigonometric polynomial (3) we use the norm in the Wiener algebra \(A({\mathbb {T}})\).
Remark 1
The NFFT with a window function of the set \(\Phi _{m,N_1}\) is simpler than the NFFT with Kaiser–Bessel or Gaussian window function \(\varphi \), since both window functions are supported on whole \({\mathbb {R}}\). For such a window function \(\varphi \), an additional step in the NFFT is necessary, where the 1-periodic function (4) is approximated by the 1-periodic well-localized function
where \({{\tilde{\psi }}}\) is the 1-periodization of the truncated window function (see [18, pp. 378–381])
Thus the NFFT with a window function \(\varphi \) supported on whole \({\mathbb {R}}\) requires also a truncated version of \(\varphi \). In this case, the error of the NFFT is measured by \(\Vert s_1 - f\Vert _{C(\mathbb T)}\). In [18, p. 393], it is shown that the error of the NFFT with the Kaiser–Bessel window function can be estimated by
We will see in Subsections 5.4 – 5.6 that special window functions \(\varphi \in \Phi _{m,N_1}\) possess a similar error behavior as the Kaiser–Bessel window function.
We say that the window function \(\varphi \in \Phi _{m,N_1}\) is convenient for NFFT, if the \(C({\mathbb {T}})\)-error constant
with
fulfills the condition \(e_{\sigma }(\varphi ) \ll 1\) for conveniently chosen truncation parameter \(m \ge 2\) and oversampling factor \(\sigma > 1\). Later in Theorem, 2 we will show that under certain assumptions on \(\varphi \in \Phi _{m,N_1}\) the value \(e_{\sigma ,N}(\varphi )\) is bounded for all \(N \in {\mathbb {N}}\). This \(C({\mathbb {T}})\)-error constant is motivated by techniques first used in [24] and later also in [5]. G. Steidl [24] has applied this technique for error estimates of NFFT with B-spline and Gaussian window functions, respectively.
Lemma 2
For each \(N\in 2 {\mathbb {N}}\), the constant \(e_{\sigma ,N}(\varphi )\) of \(\varphi \in \Phi _{m,N_1}\) can be represented in the equivalent form
Proof
By Lemma 1 we know that for all \(x\in {\mathbb {T}}\) it holds
and hence
This completes the proof. \(\square \)
Thus the condition \(e_{\sigma }(\varphi ) \ll 1\) with (7) means that each exponential
can be approximately reproduced by a linear combination of shifted window functions \({{{\tilde{\varphi }}}}(\cdot \, + \frac{\ell }{N_1})\) with \(\ell \in I_{N_1}\). In other words, the equispaced shifts \(\tilde{\varphi }( \cdot + \frac{\ell }{N_1})\) with \(\ell \in I_{N_1}\) are approximately exponential reproducing. For each node \(x_j \in [-\frac{1}{2}, \frac{1}{2})\), \(j \in I_M\), the linear combination
has only few nonzero terms, since the support of \(\varphi \) is very small for large \(N_1\). If we replace \(\exp (2\pi {\mathrm i} n x_j)\) for each \(n\in I_N\) and \(j\in I_M\) by the approximate value (8), we compute approximate values of
in the form
mainly by DFT. This is the key of the NFFT with nonequispaced spatial data and equispaced frequencies. Special window functions \(\varphi \in \Phi _{m,N_1}\) which are convenient for NFFT will be presented in Sect. 5.
Lemma 3
Let \(\sigma > 1\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), \(N \in 2 {\mathbb {N}}\), and \(N_1 =\sigma N \in 2 {\mathbb {N}}\) be given. Further let \(\varphi \in \Phi _{m,N_1}\). Let f be a 1-periodic trigonometric polynomial (3) and s its approximating 1-periodic function (4).
Then the error of \(\mathrm {NFFT}\) with nonequispaced spatial data and equispaced frequencies can be estimated by
Proof
From (5) it follows that
Note that \(c_n({{\tilde{\varphi }}}) = {{\hat{\varphi }}}(n) > 0\) for \(n \in I_N\) by assumption \(\varphi \in \Phi _{m,N_1}\). Then by Hölder’s inequality we obtain that for all \(x\in {\mathbb {T}}\) it holds
Hence we get (9). \(\square \)
Remark 2
Let \(\lambda \ge 0\) be fixed. We introduce the 1-periodic Sobolev space \(H^{\lambda }({\mathbb {T}})\) of all 1-periodic functions \(f:{\mathbb {T}} \rightarrow {\mathbb {C}}\) which are integrable on \([0,\,1]\) and for which
where we declare \(|0|:=1\). Then \(H^{\lambda }({\mathbb {T}})\) is a Hilbert space with the inner product
For \(\lambda =0\), we have \(H^{0}({\mathbb {T}})=L^{2}({\mathbb {T}})\). Then the Sobolev embedding theorem (see [23, p. 142]) says that for \(\lambda > \frac{1}{2}\) it holds \( H^{\lambda }({\mathbb {T}}) \subset A({\mathbb {T}}) \subset C({\mathbb {T}}). \) Let f be a 1-periodic, trigonometric polynomial of the form (3). Then we have
Applying the Cauchy–Schwarz inequality, we obtain for \(\lambda > \frac{1}{2}\) that
Using the Riemann zeta function \( \zeta (2\lambda ) := \sum _{k=1}^{\infty } \frac{1}{k^{2\lambda }}\), \(\lambda > \frac{1}{2}, \) we obtain the following inequality
Thus under the assumptions of Lemma 3, the error of NFFT with nonequispaced spatial data and equispaced frequencies can be estimated by
Note that
2.2 NFFT with nonequispaced frequencies and equispaced spatial data
The NFFT with nonequispaced frequencies and equispaced spatial data or transposed NFFT evaluates the exponential sums
for arbitrary given coefficients \(f_j \in {\mathbb {C}}\) and nonequispaced frequencies \(x_j\in \big [-\frac{1}{2},\,\frac{1}{2}\big )\), \(j\in I_M\). Assume that the window function \(\varphi \in \Phi _{m,N_1}\) is convenient for NFFT. Introducing the 1-periodic function
the Fourier coefficients of g read as follows
Using the trapezoidal rule, we approximate \(c_k(g)\) by
Note that \({{\widetilde{c}}}_k(g)\), \(k \in I_N\), can be efficiently computed by FFT, for details see [18, p. 382]. Then the results of this NFFT with nonequispaced frequencies and equispaced spatial data are the values
It is interesting that the same \(C({\mathbb {T}})\)-error constant (6) appears in an error estimate of the NFFT with nonequispaced frequencies and equispaced spatial data too.
Lemma 4
Let \(\sigma > 1\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), \(N,\,M \in 2 {\mathbb {N}}\), and \(N_1 = \sigma N \in 2 {\mathbb {N}}\). Further let \(\varphi \in \Phi _{m,N_1}\). For given \(f_j \in {\mathbb {C}}\) and nonequispaced frequencies \(x_j \in \big [- \frac{1}{2},\, \frac{1}{2}\big )\), \(j \in I_M\), we consider the exponential sums (10) and the related approximations (11).
Then the error of \(\mathrm {NFFT}\) with nonequispaced frequencies and equispaced spatial data can be estimated by
Proof
For each \(k \in I_N\) we have
From Hölder’s inequality and Lemma 2 it follows that
This completes the proof. \(\square \)
3 Auxiliary estimates
In our study we use later the following
Lemma 5
For \(-1< u < 1\) and \(\mu > 1\) it holds
Proof
For \(-1< u < 1\), \(r\in {\mathbb {N}}\), and \(\mu > 1\) we have
Using (12), the series can be estimated as follows
Hence it follows by the integral test for convergence that
\(\square \)
For fixed \(\mu \ge 0\), the \(\mu \) th Bessel function of first kind is defined by
so that in particular
For fixed \(\mu \ge 0\), the \(\mu \) th modified Bessel function of first kind is defined by
so that in particular
For the properties of Bessel functions we refer to [1, pp. 355–478] and [25]. In particular, these Bessel functions possess the following asymptotic behaviors for \(x\rightarrow \infty \) (see [1, pp. 364, 377]),
Here we are interested in explicit error estimates for NFFT with compactly supported, continuous window function. For this purpose, we need explicit bounds for the Bessel functions instead of the asymptotic formulas (13) and (14).
Lemma 6
For fixed \(\mu > \frac{1}{2}\) and all \(x\ge 0\), it holds
where \(\sqrt{\frac{2}{\pi }}\) is the best possible upper bound. In particular for \(\mu =1\) and \(x \ge 6\), we have
For \(\mu =\frac{5}{2}\) and \(x \ge 6\), we have
For \(\mu =3m\) with \(m \in {\mathbb {N}} {\setminus } \{1\}\) and \(x \ge \pi m\), we have
Further for all \(x \in {\mathbb {R}}\) with \(|x| \ge \pi \), it holds
Proof
The inequality (15) was shown in [13]. For \(\mu =1\), the inequality (15) means that
Thus for all \(x \ge 6\) we have
with the constant
such that
Similarly, one can show the inequalities (16) and (17).
Since \(J_0\) is even, we prove the inequality (18) only for positive x. By [25, p. 206], the Bessel function \(J_0(x)\) can be represented for \(x > 0\) in the form
where P(x) and Q(x) are certain functions as defined in [25, p. 205]. By [25, p. 208], these functions can be estimated as follows
Applying the Cauchy–Schwarz inequality to
we obtain for \(x \ge 2\),
such that
Hence it follows that for \(x \ge \pi \),
This completes the proof. \(\square \)
Lemma 7
Let \(\mu \ge \frac{1}{2}\) and \(x_0 >0\) be given. Then for all \(x\ge x_0\), it holds
Proof
By [2, Inequality (2.6)] one knows that for fixed \(\mu \ge \frac{1}{2}\) and arbitrary x, \(y\in (0,\,\infty )\) with \(x<y\) it holds
i.e.,
Hence the function \(f(x):= \sqrt{2\pi x}\,{\mathrm e}^{-x}\,I_{\mu }(x)\) is strictly increasing on \((0,\,\infty )\). Further by (14) we have
This implies the inequality (19). \(\square \)
4 Paley–Wiener theorem in a Sobolev space
The main tool of this approach is the study of the Fourier transform \({{\hat{\varphi }}}(v)\) of \(\varphi \in \Phi _{m,N_1}\) for \(|v| \rightarrow \infty \). A rapid decay of \({\hat{\varphi }}\) is essential for a small \(C({\mathbb {T}})\)-error constant (6). From Fourier analytical point of view, it is very interesting to discuss the relation to the known Theorem of Paley–Wiener (see [17, pp. 12–13]), since this result characterizes the behavior of Fourier transforms of functions \(\varphi \) which vanish outside the open interval \(I =\big (-\frac{m}{N_1},\,\frac{m}{N_1}\big )\). The smoothness of the restricted window function \(\varphi |_I\) determines the decay of \({{\hat{\varphi }}}(v)\) for \(|v| \rightarrow \infty \).
For simplicity, we denote functions defined on I by \(\varphi \) too. By \(H^k(I)\), \(k\in {{\mathbb {N}}}\), we denote the Sobolev space of all functions \(\varphi \in L^2(I)\) with \(D^j \varphi \in L^2(I)\), \(j = 1,\ldots , k\), where \(D^j \varphi \) is the jth weak derivative of \(\varphi \). Then \(H^k(I)\) is a Hilbert space with the Sobolev norm
Further we define the Sobolev space \(H_0^k(I)\) as the closure in \(H^k(I)\) of the space \(C_0^{\infty }(I)\) which consists of all infinitely differentiable, locally supported functions. Then \(H_0^k(I)\) is the set of all functions \(\varphi \in H^k(I)\) with vanishing one-sided derivatives \(\varphi ^{(j)}\big (\pm \frac{m}{N_1}\big ) = 0\), \(j =0,\ldots k-1\). We present the following modification of the Theorem of Paley–Wiener (cf. [17, pp. 12–13] or [27, Vol. II, pp. 272–274]):
Theorem 1
For given \(\varphi \in H_0^k(I)\) with \(k\in {{\mathbb {N}}}\), the function
is entire and has the following properties:
1. For real variable x, it holds \(f(x) \in L^2({\mathbb {R}})\) and \(x^k\,f(x) \in L^2({\mathbb {R}})\).
2. For all \(z \in {\mathbb {C}}\), there exist a positive constant \(\gamma _k(\varphi )\) such that
Conversely, if an entire function f satisfies the conditions 1. and 2., then the function
has the properties \(\varphi |_ I \in H^k(I)\) and \(\varphi |_{ {{\mathbb {R}}} {\setminus } I} = 0\). Note that
means the limit in \(L^2({\mathbb {R}})\) and that \(\varphi |_{ {\mathbb R} {\setminus } I} = 0\) means \(\varphi (x) = 0\) for almost all \(x\in {{\mathbb {R}}} {\setminus } I\).
Proof
Let \(\varphi \in H_0^k(I)\) be given, i.e., \(D^j\varphi \in L^2(I)\), \(j =0,\ldots , k\), and
For arbitrary \(z= x + {\mathrm i}\,y \in {\mathbb {C}}\) and \(t \in \bar{I}\) we have
and hence
Since by the Schwarz inequality
we have \(\varphi \in L^1(I)\) too. Since \(D^j\varphi \in L^2(I)\) for \(j = 1,\,\ldots , \,k\), we have \(D^j\varphi \in L^1(I)\) analogously to (22).
Obviously, f is an entire function, because for each \(z\in \mathbb C\) and \(j \in {\mathbb {N}}\) it holds
Using (20), repeated integration by parts applied to the function \(\varphi \) gives for \(j=1,\ldots , k\) the equalities
such that
Hence we obtain
From (21) and (24) it follows that for all \(z \in \mathbb C\) it holds
with the positive constant
By (23) with \(j=k\), we have for \(x \in {\mathbb {R}}\) the equalities
such that \(x^k\,f(x) \in L^2(I)\) by the Theorem of Plancherel.
Now we assume that an entire function f with the properties 1. and 2. is given. Then especially we have \(f|_{\, {\mathbb {R}}} \in L^2({\mathbb {R}})\) and
By the original Theorem of Paley–Wiener, the function
vanishes for almost all \(x \in {{\mathbb {R}}}{\setminus } I\) such that \(\varphi |_I \in L^2(I)\). For \(t\in {\mathbb {R}}\), from \(f(t) \in L^2({\mathbb {R}})\) and \(t^k\,f(t) \in L^2({\mathbb {R}})\) it follows \(t^j\,f(t) \in L^2({\mathbb {R}})\), \(j=1,\ldots ,k-1\), since \(|t^j\,f(t)|^2 \le |f(t)|^2\) for \(t\in [-1,\,1]\) and \(|t^j\,f(t)|^2 \le |t^k\,f(t)|^2\) for \(t\in {\mathbb {R}} {\setminus } [-1,\,1]\).
Each function in \(L^2({\mathbb {R}})\) generates a tempered distribution. By the differentiation property of the Fourier transform of tempered distributions we conclude that the jth weak derivative of \(\varphi \) exists almost everywhere and that
Now we have to show that \(D^j \varphi |_{{{\mathbb {R}}}{\setminus } I} = 0\). From property 2. of the entire function f it follows that the entire function \((2\pi {\mathrm i}z)^j\,f(z)\) with \(j = 1,\ldots , k\) fulfills the inequality
for all \(z\in {\mathbb {C}}\). Applying the original Theorem of Paley–Wiener to the entire function \((2 \pi {\mathrm i}z)^j\,f(z)\), we see that
vanishes almost everywhere in \({\mathbb {R}} {\setminus } I\). Since \(\psi _j(x) = (D^j \varphi )(x) =0\) almost everywhere, we have \((D^j \varphi )(x) =0\) for almost all \(x \in {\mathbb {R}} {\setminus } I\). Hence we have \(D^j \varphi |_{I} \in L^2(I)\) for \(j=1,\ldots ,k\) such that \(\varphi |_I \in H^k(I)\). This completes the proof. \(\square \)
Unfortunately, the assumption \(\varphi \in H_0^k(I)\) is too strong for the most popular window function.
Example 1
The triangular function \(\varphi (x) := 1 - \frac{N_1}{m}\,|x|\), \(x \in I\), belongs to the Sobolev space \(H_0^1(I)\), but doesn’t belong to \(H_0^2(I)\), since it holds \(\varphi \big (\pm \frac{m}{N_1}\big ) = 0\) and \(\varphi '\big (\pm \frac{m}{N_1}\big ) \ne 0\). In this case, we obtain that for \(z\in {\mathbb {C}}\),
with
Thus Theorem 1 results in \(k = 1\). Otherwise we observe a quadratic decay of the Fourier transform \({\hat{\varphi }}\), since for \(x \in {\mathbb {R}} {\setminus } \{0\}\),
Hence we will present a better method in the following Theorem 2.
5 Special window functions
In this section, we determine upper bounds of the error constant (6) for various special window functions \(\varphi \in \Phi _{m,N_1}\) by two methods. If the series
then by (6) we have that
This technique can be applied for (modified) B-spline window functions.
For the algebraic, Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window functions, we use the following argument. By Hölder’s inequality it follows from (6) that
Thus we show that the minimum of all \({{\hat{\varphi }}}(n)\) for relatively low frequencies \(n\in I_N\) is equal to \({{\hat{\varphi }}}(N/2)\) and that the series
is bounded for each \(n\in I_N\). For this we have to estimate the Fourier transform \({{\hat{\varphi }}}(n)\) for sufficiently large frequencies \(|n| \ge N_1 -\frac{N}{2}\) very carefully. Thus this results in
Now we use the special structure of the given window function \(\varphi \). Let \(m \in {{\mathbb {N}}}{\setminus } \{1\}\) and \(\sigma > 1\) be given. Assume that an even, continuous function \(\varphi _0: {{\mathbb {R}}} \rightarrow [0,\,1]\) with \(\mathrm {supp}\,\varphi _0 = [-1,\,1]\) has following properties: \(\varphi _0(0) = 1\), \(\varphi _0|_{[0,\,1]}\) is decreasing, and its restricted Fourier transform \({\hat{\varphi }}_0|_{[0, \,m/(2\sigma )]}\) is positive and decreasing. Let \(N \in 2{{\mathbb {N}}}\) with \(N_1 = \sigma N \in 2{{\mathbb {N}}}\) be given. Then the scaled function
is a continuous window function of the set \(\Phi _{m,N_1}\).
Theorem 2
Let \(\sigma > 1\), \(m \in \mathbb N {\setminus } \{1\}\), \(N \in 2 {\mathbb {N}}\), and \(N_1 = \sigma N \in 2 {\mathbb {N}}\) be given. Further let \(\varphi \in \Phi _{m,N_1}\) be a scaled version (26) of \(\varphi _0\). Assume that the Fourier transform \({{\hat{\varphi }}}_0\) fulfills the decay condition
with certain constants \(c_1 > 0\), \(c_2>0\), and \(\mu > 1\).
Then the constant \(e_{\sigma ,N}(\varphi )\) is bounded for all \(N \in 2{{\mathbb {N}}}\) by
Further, the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi )\) of the window function (26) has the upper bound
Proof
Note that it holds
By the scaling property of the Fourier transform, we have
For all \(n \in I_N\) and \(r\in {{\mathbb {Z}}}{\setminus } \{0,\,\pm 1\}\), we obtain
and hence
From Lemma 5 it follows that for fixed \(u =\frac{n}{N_1}\in \big [- \frac{1}{2\sigma },\, \frac{1}{2\sigma }\big ]\),
For all \(n \in I_N\), we sustain
since
Thus we estimate for each \(n \in I_N\),
Now we determine the minimum of all positive values
Since \(\frac{m\,|n|}{N_1} \le \frac{m}{2 \sigma }\) for all \(n \in I_N\), we obtain
Thus we see that the constant \(e_{\sigma ,N}(\varphi )\) can be estimated by an upper bound which depends on m and \(\sigma \), but does not depend on N. We obtain
Consequently, the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi )\) has the upper bound (27). \(\square \)
In the following, we show that for our study, Theorem 2 is more convenient than Theorem 1 (cf. Example 1).
Example 2
For \(m =2\) and \(\sigma > 1\), we consider the triangular window function (26) with
Then the Fourier transform of \(\varphi _0\) reads as follows
Obviously, \({\varphi }_0\) is an even, continuous function with \(\mathrm {supp}\,\varphi _0 = [-1,\,1]\), where \(\varphi _0(0) = 1\) and \(\varphi _0|_{[0,\,1]}\) is decreasing. The restricted Fourier transform \({{\hat{\varphi }}}_0|_{[0,\,1/\sigma ]}\) is positive and decreasing. Further \({{\hat{\varphi }}}_0\) fulfills the decay condition
Using Theorem 2, we obtain the following upper bound of the \(C({\mathbb {T}})\)-error constant
Note that a better result will be presented in Theorem 3 (for \(m = 1\)).
5.1 B-spline window function
We start with the popular B-spline window function (see [6, 24]). Assume that \(N\in 2\,{\mathbb {N}}\) and \(\sigma > 1\) with \(N_1 = \sigma N \in 2\,{\mathbb {N}}\) are given. We consider the \(\mathrm {B}\)-spline window function
where \(M_{2m}\) denotes the centered cardinal B-spline of even order 2m with \(m\in {\mathbb {N}}\). For \(m=1\), we obtain the triangular window function. Using the three-term recursion
with
we find
Note that \(M_{2m}(0) > 0\) for all \(m \in {\mathbb {N}}\), since it holds \(M_{2m}(x) > 0\) for each \(x \in (-m,\,m)\). As known, the Fourier transform of (28) (see [18, p. 452]) has the form
If \({{{\tilde{\varphi }}}}_{\mathrm B}\) is the 1-periodization (1) of \(\varphi _{\mathrm B}\), then the Fourier coefficients of \({{\tilde{\varphi }}}_{\mathrm B}\) read as follows
Note that \(c_k({{\tilde{\varphi }}}_{\mathrm B}) > 0\) for all \(k \in I_N\).
By (5) we see that
Now we estimate the \(C({\mathbb {T}})\)-error constants for NFFT. Applying the special structure of the Fourier coefficients \(c_k({{{\tilde{\varphi }}}}_{\mathrm B})\) and Lemma 5, we obtain a good upper bound (25) of the \(C(\mathbb T)\)-error constant by this method. For a proof of the following result see [24].
Theorem 3
Let \(\sigma > 1\), \(N\in 2\,{\mathbb {N}}\), and \(N_1 = \sigma N \in 2\,{\mathbb {N}}\). Further \(m \in {\mathbb {N}}\) with \(2m \ll N_1\) is given. Then the \(C({\mathbb {T}})\)-error constant (6) of the \(\mathrm {B}\)-spline window function (28) can be estimated by
i.e., the \(\mathrm {B}\)-spline window function (28) is convenient for \(\mathrm {NFFT}\).
5.2 Modified B-spline window function
Let \(\sigma \ge 1\), \(m\in {\mathbb {N}} {\setminus } \{1\}\), \(N\in 2\,{\mathbb {N}}\), and \(N_1 = \sigma N\) be given. The approach to the B-spline window function (28) can be generalized to the modified \(\mathrm {B}\)-spline window function (see [14])
where \(M_{2b}\) denotes the centered cardinal B-spline of order \(2b\in {\mathbb {N}} {\setminus } \{1, \, 2\}\), i.e., \(b \in \{\frac{3}{2},\,2,\,\frac{5}{2},\,\ldots \}\). Assume that \(N_1 b \in 2\,{\mathbb {N}}\) and that \(m \in {\mathbb {N}}\) fulfills the conditions \(m < 2\sigma b\) and \(b\not = m\). Using the three-term recursion (29), we find \( M_3(0) =\frac{3}{4}\), \(M_5(0) = \frac{115}{192}. \) Obviously, it holds \(\varphi _{\mathrm {mB}} \in \Phi _{m, N_1}\), where the Fourier transform of (30) (see [18, p. 452]) reads as follows
If \({{\tilde{\varphi }}}_{\mathrm {mB}}\) denotes the 1-periodization of (30), then the Fourier coefficients of \(\tilde{\varphi }_{\mathrm {mB}}\) have the following form
Note that \(c_k({{{\tilde{\varphi }}}}_{\mathrm {mB}}) > 0\) for all \(k \in I_N\). Let f be an arbitrary 1-periodic trigonometric polynomial (3) which we approximate by the 1-periodic function
with conveniently chosen coefficients \(g_{\ell } \in {\mathbb {C}}\). Then s possesses the Fourier expansion
with the Fourier coefficients
where
Thus the vector \(({{\hat{g}}}_k)_{k^\in I_{N_1 b}}\) is equal to the DFT of length \(N_1 b\) of the vector \((g_{\ell })_{\ell \in I_{N_1 b}}\) and we have \({{\hat{g}}}_{k + r N_1 b} = {{\hat{g}}}_k\) for all \(k\in I_{N_1 b}\) and \(r\in {\mathbb {Z}}\). In order to approximate f by s, we choose
Thus we see that \(c_k(s) = c_k(f)\) for all \(k \in I_N\) and \(c_k(s) = 0\) for all \(k \in I_{N_1 b} {\setminus } I_N\). Substituting \(k = n + r N_1 b\) with \(n\in I_{N_1 b}\) and \(r \in {{\mathbb {Z}}}\), we obtain
such that
Then it follows by Hölder’s inequality that
Theorem 4
Let \(\sigma \ge 1\), \(N\in 2\,{\mathbb {N}}\), \(2 b \in {\mathbb {N}} {\setminus } \{1,\,2\}\), and \(N_1 b\in 2\,{\mathbb {N}}\) with \(N_1 = \sigma N\). Further let \(m \in {\mathbb {N}}\) with \(2 b > m\) and \(b\not = m\) be given.
Then the \(C({\mathbb {T}})\)-error constant of \(\mathrm {NFFT}\) with the modified \(\mathrm {B}\)-spline window function (30) can be estimated by
i.e., the modified \(\mathrm {B}\)-spline window function (30) is convenient for \(\mathrm {NFFT}\).
Proof
Now we estimate
For \(n=0\) and \(r \in {\mathbb {Z}} {\setminus } \{0\}\) we have
For \(n \in I_N {\setminus } \{0\}\) and \(r \in {\mathbb {Z}} {\setminus } \{0\}\) we obtain
since for \(n \in I_N\) it holds
Using Lemma 5, we conclude that
with \(2 \sigma b \ge 3\). \(\square \)
5.3 Algebraic window function
For fixed shape parameter \(\beta = 3 m\) with \(m \in {\mathbb {N}} {\setminus } \{1\}\) and for an oversampling factor \(\sigma > \frac{\pi }{3}\), we consider the algebraic window function
Theorem 5
Let \(N \in 2 {\mathbb {N}}\) and \(\sigma > \frac{\pi }{3}\), where \(N_1 = \sigma N \in 2 {\mathbb {N}}\). Further let \(m \in {\mathbb {N}} {\setminus } \{1\}\) and \(\beta = 3m\) be given.
Then the \(C({\mathbb {T}})\)-error constant of the algebraic window function (31) can be estimated by
i.e., the algebraic window function (31) is convenient for \({\mathrm {NFFT}}\).
Proof
We apply Theorem 2 and introduce the unscaled algebraic window function
Using [16, p. 8], we determine the corresponding Fourier transform
Thus \({{\hat{\varphi }}}_{0, \mathrm {alg}}(v)\) for \(v>0\) is a multiple of the function \((\pi v)^{-\beta }\,J_{\beta }(2 \pi v)\) which can be represented as infinite product
where \(j_{\beta ,s}\) denotes the sth positive zero of \(J_{\beta }\) (see [1, p. 370]). Note that
For \(\beta = 3m\) it holds \(j_{\beta ,1} > 3m + \pi - \frac{1}{2}\) (see [10]). Hence by \(\sigma > \frac{\pi }{3}\) we get
Thus each factor of the infinite product (34) is positive and decreasing for \(v \in \big [0,\, \frac{m}{2\sigma }\big ]\). Hence by (33) and (34), the Fourier transform \({{\hat{\varphi }}}_{0, \mathrm {alg}}(v)\) is positive and decreasing for \(v \in \big [0,\, \frac{m}{2\sigma }\big ]\) and it holds
By (17) we know that for \(|v|\ge \frac{m}{2}\) it holds
For \(|v| \ge m\,\big (1 - \frac{1}{2 \sigma }\big )\) and \(\sigma > \frac{\pi }{3}\) it holds
Using (33) and (36), we obtain for \(|v| \ge m\,\big (1 - \frac{1}{2 \sigma }\big )\) the estimate
Applying Theorem 2, we obtain
where by (37) we have \(\mu = \beta + \frac{1}{2} = 3m + \frac{1}{2}\) and
Using (35), we get the inequality (32). \(\square \)
5.4 Bessel window function
Let \(N \in 2\, {\mathbb {N}}\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) be given. For fixed shape parameter
we consider the new Bessel window function
with \(N_1 = \sigma N \in 2 {\mathbb {N}}\). Obviously, the Bessel window function is continuously differentiable and compactly supported.
Theorem 6
Let \(N \in 2 {\mathbb {N}}\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) be given, where \(N_1 = \sigma N \in 2 {\mathbb {N}}\) and \(2m \ll N_1\).
Then the \(C({\mathbb {T}})\)-error constant of the Bessel window function (39) with the shape parameter (38) can be estimated by
i.e., the Bessel window function (39) is convenient for \({\mathrm {NFFT}}\).
Proof
We apply Theorem 2 and introduce the unscaled Bessel window function
We determine the even Fourier transform
By [16, p. 96], this Fourier transform reads as follows
Introducing the spherical Bessel function (see [1, pp. 437–438])
and the modified spherical Bessel function (see [1, p. 443])
we obtain
Using the power series expansion of the modified spherical Bessel function \(i_2\) (see [1, p. 443]), for \(|v| < m\,\big (1 - \frac{1}{2 \sigma }\big )\) we receive
Hence \({{\hat{\varphi }}}_{0,\mathrm {Bessel}}(v)\) is positive and decreasing for \(v \in \big [0,\, m \big (1 - \frac{1}{2\sigma }\big ) \big )\). Since \(\frac{m}{2 \sigma } < m \big (1 - \frac{1}{2 \sigma }\big )\) for \( \sigma \ge \frac{5}{4}\), we conclude
For \(m\ge 2\) and \(\sigma \ge \frac{5}{4}\), we maintain
Applying the inequality (19) with \(\mu = \frac{5}{2}\), for \(x \ge x_0\) we obtain
Hence for \(x = 2 \pi m \sqrt{1 - \frac{1}{\sigma }}\) it follows that
Thus we see that
Now we estimate the Fourier transform \({\hat{\varphi }}_{0,\mathrm {Bessel}}(v)\) for \(|v| \ge m\,\big (1 + \frac{1}{2 \sigma }\big )\). Using the assumptions \(m \ge 2\) and \(\sigma \in \big [\frac{5}{4},\,2\big ]\), for \(|v| \ge m\,\big (1 + \frac{1}{2 \sigma }\big )\) we get
By (16), for all \(x \ge 6\) it holds
Thus for \(|v| \ge m\, \big (1 + \frac{1}{2 \sigma }\big )\), we obtain by (40) that
Since the function \(g:\,\big [m\,\big (1 + \frac{1}{2\sigma }\big ),\, \infty \big ) \rightarrow {\mathbb {R}}\) defined by
is decreasing and bounded from above by
we receive for \(|v| \ge m\, \big (1 + \frac{1}{2 \sigma }\big )\) the estimate
Finally we show that for \(|v| \ge m \,\big (1 - \frac{1}{2 \sigma }\big )\) it holds
By (41) we have
For \(|v| > m \big (1 - \frac{1}{2 \sigma }\big )\), it holds by (41) that
By the definition of the spherical Bessel function \(j_2\) it holds for \(x> 0\) that
Then from [25, p. 49] it follows that for \(x > 0\),
with \(\Gamma (7/2) = \frac{15}{8}\,\sqrt{\pi }\) such that by \(j_2(0) = 0\) we obtain the inequality
and hence (44).
Applying (27) together with (42), (43), and (44), we conclude that
with
For \(\sigma \in \big [\frac{5}{4}, \,2\big ]\), it holds
This completes the proof. \(\square \)
5.5 sinh-type and related window functions
Let \(N \in 2 {\mathbb {N}}\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) be given. For fixed shape parameter \(\beta \) as stated in (38), we consider the new \(\sinh \)-type window function
with \(N_1 = \sigma N \in 2 {\mathbb {N}}\). This \(\sinh \)-type window function belongs to \(\Phi _{m,N_1}\). Note that (45) is not piecewise continuously differentiable, since \(\varphi _{\sinh }'(-\frac{m}{N_1}+0)=\infty \) and \(\varphi _{\sinh }'(\frac{m}{N_1}-0)= -\infty \). Note that the \(\sinh \)-type window function (45) can be computed much faster than the Bessel window function (39).
Theorem 7
Let \(N \in 2 {\mathbb {N}}\), \(N \ge 8\), \(m \in {\mathbb {N}} {\setminus } \{1\}\), and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) be given, where \(N_1 = \sigma N \in 2 {\mathbb {N}}\) and \(2m \ll N_1\). The shape parameter \(\beta \) is given by (38).
Then the \(C({\mathbb {T}})\)-error constant of the \(\sinh \)-type window function (45) with the shape parameter (38) can be estimated by
i.e., the \(\sinh \)-type window function (45) is convenient for \({\mathrm {NFFT}}\).
A proof of Theorem 7 can be found in [21]. The proof based mainly on the knowledge of the analytical Fourier transform. In [21] we consider in addition two related window functions, namely the continuous \(\exp \)-type window function
as well as the continuous \(\cosh \)-type window function
The main drawback for the numerical analysis of the \(\exp \)-type/\(\cosh \)-type window function is the fact that an explicit Fourier transform of this window function is unknown. Therefore we split the continuous \(\exp \)-type/\(\cosh \)-type window function into a sum \(\psi + \rho \), where the Fourier transform of the compactly supported function \(\psi \) is explicitly known and where the compactly supported function \(\rho \) has small magnitude. For details see [21]. Note that \(\varphi _{\mathrm {exp}}\) was first suggested in [2, 4] and a discontinuous version of \(\varphi _{\cosh }\) was suggested in [4, Remark 13].
5.6 Modified cosh-type and related window functions
For \(\sigma \in \big [\frac{5}{4},\,2 \big ]\) and \(m \in {\mathbb {N}} {\setminus } \{1\}\), we consider the modified \(\cosh \)-type window function
where it holds
with the shape parameter \(\beta \) as stated in (38). By [16, p. 6 and p. 38], its Fourier transform reads as follows
Note that the Fourier integral in [16, p. 38, formula 7.57] reads as follows
Obviously, the unscaled modified \(\cosh \)-type window function \({\varphi }_{0,\mathrm {mcosh}}:\,{\mathbb {R}} \rightarrow [0,\,1]\) with the support \([-1,\,1]\) is even and continuous on \({\mathbb {R}}\). Further the restricted function \({\varphi }_{0,\mathrm {mcosh}}|_{[0,\,1]}\) is decreasing. Now we prove that the Fourier transform (47) is positive and decreasing in \(\big [0,\,\frac{\beta }{2\pi }\big ]\). First we remark that by (47) it holds
since \(J_0(0) = 1\) and \(|J_0(\beta )| < \frac{1}{2}\) for \(\beta = 2\pi m \big (1 - \frac{1}{2 \sigma }\big ) \ge \frac{12\, \pi }{5}\). Using \(I_0'(x) = I_1(x)\) and \(J_0'(x) = - J_1(x)\), from (47) it follows for all \(v \in \big (0,\,\frac{\beta }{2 \pi }\big )\) that
Since by Lemma 8 it holds the inequality
the Fourier transform (47) is decreasing in \(\big [0,\,\frac{\beta }{2 \pi }\big ]\).
Lemma 8
Let \(\beta \) be given by (38) with \(m \in {{\mathbb {N}}}{\setminus } \{1\}\) and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) such that \(\beta >2\). Then for \(x \in [0,\,\beta ]\) it holds
Proof
Obviously, the inequality (48) holds for \(x = 0\) and \(x = \beta \), since \(J_1(0) = I_1(0) = 0\). First we prove (48) for \(x \in \big (0,\,\frac{\beta }{\sqrt{2}}\big ]\). By the known power expansions
we obtain for \(x \in \big (0, \,\frac{\beta }{\sqrt{2}}\big ]\) that
Then from \(x^2 \le \beta ^2 - x^2\) for \(x \in \big (0, \,\frac{\beta }{\sqrt{2}}\big ]\) it follows by (49) and (50) that
i.e., this implies the inequality (48) for \(x \in \big (0, \,\frac{\beta }{\sqrt{2}}\big ]\).
In the case \(x \in \big [\frac{\beta }{\sqrt{2}}\,, \,\beta \big )\), we substitute \(y:=\sqrt{\beta ^2 - x^2} \in \big (0,\,\frac{\beta }{\sqrt{2}}\big ]\). Then we show that
This inequality is fulfilled, since by [25, p. 31],
and hence
since it holds \(\beta > 2\) as well as \(\frac{y}{2} \le I_1(y)\) for \(y \ge 0\) by \(I_1(0) = 0\), \(I_1'(0) = \frac{1}{2}\) and \(I_1''(y) > 0\). This completes the proof. \(\square \)
Thus we obtain
From \(m \ge 2\) and \(\sigma \in \big [\frac{5}{4},\,2\big ]\) it follows that
Hence by (18) it holds
This implies that
Further for \(v \in \big [m \big (1 - \frac{1}{2 \sigma }\big ),\,m \big (1 + \frac{1}{2 \sigma }\big )\big ]\) we obtain
since by [25, p. 31] and (18) it holds
Thus we can use \(c_1 =\frac{3 \pi }{2\,(\cosh \beta - 1)}\) as constant in Theorem 2.
Finally we determine the decay of (47) for \(|v| \ge m \big (1 + \frac{1}{2 \sigma }\big )\). We show that
By (47) we know that
Using the power expansion of the Bessel function \(J_0\), we obtain
Since for \(k\in {\mathbb {N}}\) it holds
we estimate
Since by (18) it holds \(|J_0(2 \pi v)| < \frac{1}{2}\) for \(|v| \ge m\,\big (1 + \frac{1}{2 \sigma }\big )\), we receive the inequality (51). Thus
are the corresponding constants of Theorem 2.
For the modified \(\cosh \)-type window function (46), we estimate the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi _{\mathrm {mcosh}})\).
Theorem 8
Let \(\sigma \in \big [\frac{5}{4},\,2\big ]\), \(m \in {{\mathbb {N}}}{\setminus } \{1\}\), and \(N_1 = \sigma _1 N\in 2 {\mathbb {N}}\) with \(2 m \ll N_1\).
Then the \(C({\mathbb {T}})\)-error constant \(e_{\sigma }(\varphi _{\mathrm {mcosh}})\) has the upper bound
Proof
From Theorem 2 with \(\mu = 2\) it follows that
Since
and
we obtain the estimate (52), since by \(\sigma \in \big [\frac{5}{4},\,2\big ]\) it holds
By [26] the modified Bessel function \(I_0(x)\) fulfills the inequalities
If we apply this result, then we obtain
This completes the proof. \(\square \)
Finally in our numerical examples, we consider also the modified \(\mathrm {exp}\)-type window function
where it holds
and the modified \(\sinh \)-type window function
where it holds
Note that the main idea to consider these modified window functions, comes from the Fourier transform of the Kaiser–Bessel window function, see [19, Remark 1]. The modified \(\sinh \)-type window function \(\varphi _{\mathrm {msinh}}\) is used in the NFFT package [11], and gives very good error bounds.
6 Conclusion
In this paper, we present explicit error estimates for the NFFT with continuous, compactly supported window function \(\varphi \in \Phi _{m,N_1}\). Such window functions simplify the algorithm for NFFT, since the truncation error of the NFFT vanishes. Using the \(C({\mathbb {T}})\)-error constant of \(\varphi \), we propose a unified approach to error estimates of the NFFT with nonequispaced spatial data and equispaced frequencies as well as of the NFFT with nonequispaced frequencies and equispaced spatial data. Further we discuss the connection with a modified Paley–Wiener theorem.
We present two techniques to find upper bounds of the \(C(\mathbb T)\)-error constant. The second method which uses the scaling structure of the window function \(\varphi \), shows that the constant \(e_{\sigma ,N}(\varphi )\) is bounded for all \(N \in 2 {\mathbb {N}}\). We see that \(e_{\sigma }(\varphi )\) depends essentially on the decay of the Fourier transform \({{\hat{\varphi }}}(v)\) for \(|v| \rightarrow \infty \) and the positive size of \({{\hat{\varphi }}}(v)\) for small frequencies. For the (modified) B-spline, algebraic, Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window functions, we present new explicit upper bounds of the corresponding \(C({\mathbb {T}})\)-error constants. Here we use the fact that the Fourier transforms of these window functions are explicitly known. It is remarkable that the \(C({\mathbb {T}})\)-error constants of Bessel, \(\sinh \)-type, and modified \(\cosh \)-type window function decay exponentially with respect to m. Numerical experiments verify the different behavior of the \(C({\mathbb {T}})\)-error constants for these window functions, see Fig. 1. We point out that the modified \(\cosh \)-type, \(\exp \)-type, and \(\sinh \)-type window functions give the best error constants, see Fig. 2. Note that in Fig. 2, the constants \(e_{\sigma ,1024}(\varphi )\) for \(\varphi \in \{\varphi _{\sinh },\,\varphi _{cosh},\,\varphi _{\exp }\}\) are numerically not distinguishable, but not identical. The same is true for \(\varphi \in \{\varphi _{\mathrm {msinh}},\,\varphi _{\mathrm {mcosh}},\,\varphi _{\mathrm {mexp}}\}\). The modified \(\sinh \)-type window function, see Sect. 5.6, is used in the NFFT package [11].
References
Abramowitz, M., Stegun, I.A. (eds.): Handbook of Mathematical Functions. National Bureau of Standards, Washington, DC (1972)
Baricz, Á.: Bounds for modified Bessel functions of the first and second kinds. Proc. Edinb. Math. Soc. (2), 53(3), 575–599 (2010)
Barnett, A.H., Magland, J.F., Klinteberg., L.A.: Flatiron Institute nonuniform fast Fourier transform libraries (FINUFFT). http://gitub.com/flatironinstitute/finufft
Barnett, A.H., Magland, J.F., Klinteberg, L.A.: A parallel non-uniform fast Fourier transform library based on an “exponential of semicircle’’ kernel. SIAM J. Sci. Comput. 41, C479–C504 (2019)
Barnett, A.: Aliasing error of the \(\exp (\beta \sqrt{1-z^2})\) kernel in the nonuniform fast Fourier transform. Appl. Comput. Harmon. Anal. 51, 1–16 (2021)
Beylkin, G.: On the fast Fourier transform of functions with singularities. Comput. Harmon. Anal. 2, 363–381 (1995)
Deserno, M., Holm, C.: How to mesh up Ewald sums. I. A theoretical and numerical comparison of various particle mesh routines. J. Chem. Phys. 109, 7678–7693 (1998)
Dutt, A., Rokhlin, V.: Fast Fourier transforms for nonequispaced data. SIAM J. Sci. Stat. Comput. 14, 1368–1393 (1993)
Dutt, A., Rokhlin, V.: Fast Fourier transforms for nonequispaced data II. Appl. Comput. Harmon. Anal. 2, 85–100 (1995)
Ifantis, E.K., Siafarikas, P.D.: A differential equation for the zeros of Bessel functions. Appl. Anal. 20, 269–281 (1985)
Keiner, J., Kunis, S., Potts, D.: NFFT 3.5, C subroutine library. http://www.tu-chemnitz.de/~potts/nfft. Contributors: F. Bartel, M. Fenn, T. Görner, M. Kircheis, T. Knopp, M. Quellmalz, M. Schmischke, T. Volkmer, A. Vollrath
Keiner, J., Kunis, S., Potts, D.: Using NFFT3 - a software library for various nonequispaced fast Fourier transforms. ACM Trans. Math. Softw. 36, Article 19, 1–30 (2009)
Krasikov, I.: Approximations for the Bessel and Airy functions with an explicit error term. LMS J. Comput. Math. 17(1), 209–225 (2014)
Nestler, F.: Automated parameter tuning based on RMS errors for nonequispaced FFTs. Adv. Comput. Math. 42(4), 889–919 (2016)
Nestler, F.: Efficient Computation of Electrostatic Interactions in Particle Systems Based on Nonequispaced Fast Fourier Transforms. Dissertation. Universitätsverlag Chemnitz (2018)
Oberhettinger, F.: Tables of Fourier Transforms and Fourier Transforms of Distributions. Springer, Berlin (1990)
Paley, R.E.A.C., Wiener, N.: Fourier Transforms in Complex Domains. AMS Colloq. Publ, Providence (1934)
Plonka, G., Potts, D., Steidl, G., Tasche, M.: Numerical Fourier Analysis. Birkhäuser/Springer, Cham (2018)
Potts, D., Steidl, G.: Fast summation at nonequispaced knots by NFFTs. SIAM J. Sci. Comput. 24, 2013–2037 (2003)
Potts, D., Steidl, G., Tasche, M.: Fast Fourier transforms for nonequispaced data: A tutorial. In: Benedetto, J.J., Ferreira, P.J.S.G. (eds.) Modern Sampling Theory: Mathematics and Applications, pp. 247–270. Birkhäuser, Boston (2001)
Potts, D., Tasche, M.: Continuous window functions for NFFT. Adv. Comput. Math. 47, 53 (2021)
Ruiz-Antolin, D., Townsend, A.: A nonuniform fast Fourier transform based on low rank approximation. SIAM J. Sci. Comput. 40(1), A529–A547 (2018)
Saranen, J., Vainikko, G.: Periodic Integral and Pseudodifferential Equations with Numerical Approximation. Springer, Berlin (2002)
Steidl, G.: A note on fast Fourier transforms for nonequispaced grids. Adv. Comput. Math. 9, 337–353 (1998)
Watson, G.N.: A Treatise on the Theory of Bessel Function. Cambridge Univ. Press, Cambridge (1995)
Yang, Z.-H., Chu, Y.-M.: On approximating the modified Bessel function of the first kind and Toader–Qi mean. J. Inequal. Appl. Paper No. 10 (2016)
Zygmund, A.: Trigonometric Series, Vol. I and II. Cambridge Univ. Press, Cambridge (2002)
Acknowledgements
The authors would like to thank Jürgen Rossmann and Winfried Sickel for several fruitful discussions on the topic. Further, the authors would like to thank the reviewers for constructive advices to improve the representation of the paper. The first author acknowledges funding by Deutsche Forschungsgemeinschaft (German Research Foundation)—Project–ID 416228727 – SFB 1410.
Open Access
This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Funding
Open Access funding enabled and organized by Projekt DEAL.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Gerlind Plonka.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Potts, D., Tasche, M. Uniform error estimates for nonequispaced fast Fourier transforms. Sampl. Theory Signal Process. Data Anal. 19, 17 (2021). https://doi.org/10.1007/s43670-021-00017-z
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s43670-021-00017-z
Keywords
- Nonequispaced fast Fourier transform
- NFFT
- Error estimate
- Wiener algebra
- Oversampling factor
- Truncation parameter
- Compactly supported window function
- Paley–Wiener theorem